Research map / Graph learning / Graph self-supervised learning
NTSFormer: A Self-Teaching Graph Transformer for Multimodal Isolated Cold-Start Node Classification
Jun Hu, Yufei He, Yuan Li, Bryan Hooi, Bingsheng He
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use self-teaching Graph Transformer with cold-start mask instead of MLP-student distillation | NTSFormer · Goodreads-NC (All test nodes) | accuracy (%): 55.65 -> 61.58 (SGKD vs NTSFormer) | Improved |
| add self-teaching instead of separate GLNN-style MLP student (ablation) | NTSFormer · Movies, Ele-fashion, Goodreads-NC | accuracy (%) | Improved |
| add multimodal graph pre-computation with separate text and visual tokens (ablation) | NTSFormer · Movies, Ele-fashion, Goodreads-NC | accuracy (%) | Improved |
| add MoE input projection instead of shared linear projection (ablation) | NTSFormer · Movies, Ele-fashion, Goodreads-NC | accuracy (%) | Improved |
| use fewer than 3 routed experts (ablation) | NTSFormer MoE input projection · Movies, Ele-fashion, Goodreads-NC | accuracy (%) | Hurt |
| use multiple Transformer layers instead of one (ablation) | NTSFormer · Movies | accuracy (%): 45.22 -> 46.12 | Improved |
| use multiple Transformer layers instead of one (ablation) | NTSFormer · Ele-fashion | accuracy (%): 83.54 -> 83.37 (2 layers); 83.67 (4 layers) | No clear effect |
| increase Transformer layers from 2 to 4 (ablation) | NTSFormer · Movies | accuracy (%): 46.12 -> 45.32 | No clear effect |
| use GNNs and multimodal GNNs under isolation | GraphSAGE, MMGCN, MGAT, MIG · Movies, Ele-fashion, Goodreads-NC | accuracy (%): MLP 55.56 vs GraphSAGE 41.65 (Goodreads-NC All) | Hurt |
| use Graph Transformers (Polynormer, SGFormer, NAGphormer) instead of MLP | isolated cold-start inference · Goodreads-NC (All test nodes) | accuracy (%): 55.56 -> 38.79 (Polynormer) | Hurt |
| distill GNN teacher into MLP student (GLNN, SGKD, SimMLP) | MLP · Movies (All test nodes) | accuracy (%): 41.85 -> 43.45 (SimMLP) | Improved |
| one-time pre-computation of graph structure for training | NTSFormer · Goodreads-NC | training time (s): 1437s (MUSE) -> 260s | Improved |
Closest papers by meaning
- VCR-Graphormer: A Mini-batch Graph Transformer via Virtual Connections (ICLR 2024 · similarity 0.87)
- UniGTE: Unified Graph–Text Encoding for Zero-Shot Generalization across Graph Tasks and Domains (NeurIPS 2025 · similarity 0.87)
- Mosaic of Modalities: A Comprehensive Benchmark for Multimodal Graph Learning (CVPR 2025 · similarity 0.87)
- What Improves the Generalization of Graph Transformers? A Theoretical Dive into the Self-attention and Positional Encoding (ICML 2024 · similarity 0.86)
- InstructG2I: Synthesizing Images from Multimodal Attributed Graphs (NeurIPS 2024 · similarity 0.86)
- Equivalence is All: A Unified View for Self-supervised Graph Learning (ICML 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
